Modularized integration-based deployment method of mobile fruit wine fermentation factory

Abstract

The invention discloses a deployment method of a mobile fruit wine fermentation factory based on modular integration. The deployment method comprises the following steps: S1, dividing fermentation modules; s2, introducing the fruit juice raw material into a first fermentation module, and outputting a first fermentation liquid in combination with a random forest and a support vector machine algorithm; s3, introducing the first fermentation liquor into a second fermentation module, constructing a dynamic model of the fermentation process by using a graph convolutional network and a generative adversarial network, and outputting second fermentation liquor; s4, introducing the second fermentation liquid into a third fermentation module, and obtaining finished fruit wine by utilizing a space-time diagram neural network and combining with a variational automatic encoder; s5, constructing a global data set; s6, integrating a multi-layer coordination algorithm by the central control system, and optimizing the overall control of the fermentation process; s7, carrying out cleaning and disinfection; and S8, disassembling the fermentation modules, and moving the fermentation modules to a target location through a transport tool to complete reassembly and connection of the fermentation modules. Intelligent control over fruit wine fermentation is achieved by means of modular design and an artificial intelligence algorithm.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fruit wine fermentation, and particularly relates to a deployment method of a mobile fruit wine fermentation factory based on modular integration. BACKGROUND

[0002] At present, fruit wine fermentation process plays an important role in the food processing industry. Traditional fruit wine fermentation factories mainly use fixed equipment for production. Such fixed fermentation factories usually include multiple links such as raw material processing, fermentation, aging and bottling. The fermentation process mostly relies on manual operation and monitoring, and relies on the experience and manual adjustment of the operator to control the fermentation environment, such as temperature, humidity, oxygen concentration, etc. However, traditional fixed fermentation factories have many limitations, mainly in flexibility, fine control, fermentation efficiency, product quality and market response, etc.

[0003] Firstly, the fixed fermentation factory lacks flexibility and is difficult to adapt to the needs of different production locations and raw material producing areas. Fruit wine production has high requirements for the freshness and producing area of raw materials, but traditional factories are mostly fixed, and raw materials need to be transported to the factory for fermentation, which often leads to a decrease in the freshness of raw materials and an increase in transportation costs. In addition, traditional fixed factories also lack flexibility in production scale, process parameter adjustment, etc., and are difficult to quickly respond to changes in market demand, and cannot realize flexible deployment of production locations.

[0004] Secondly, the traditional fermentation process is difficult to realize fine and automatic control, and the environmental parameters in the fermentation process mostly rely on manual monitoring and adjustment. Manual operation not only increases labor costs, but also due to the uncontrollability of human factors, it is easy to cause fluctuations in the fermentation process, thereby affecting the quality and flavor of fruit wine. Traditional process is difficult to accurately monitor and control key parameters in the fermentation process, especially in the key stages of sugar to alcohol conversion and flavor substance formation, it is difficult to realize real-time optimization and fine management of the fermentation process.

[0005] In addition, the traditional fruit wine fermentation process usually adopts a single fermentation process, which is difficult to fine manage according to different stages of the fermentation process. In the fermentation process, the processes of sugar conversion, alcohol generation and flavor formation have different optimal conditions, and a single fermentation process cannot meet the needs of each stage, often leading to problems such as too long fermentation time, too much by-product generation, unstable flavor, etc., affecting the quality and production efficiency of the product. At the same time, the traditional process is insufficient in monitoring and controlling by-products in the fermentation process, which may lead to abnormal generation of acidity, esters, phenols and other substances, further affecting the quality and stability of fruit wine.

[0006] In addition, the existing fruit wine fermentation process lags behind in data collection and analysis. In the traditional process, data collection usually relies on manual recording, and data analysis is mainly based on experience and judgment, lacking scientific analysis models. Such a way not only makes it difficult to realize real-time optimization of the fermentation process, but also makes it difficult to effectively utilize a large amount of historical data and form a data-driven fermentation process optimization strategy. Especially in the fermentation process, real-time monitoring and dynamic regulation of parameters such as temperature, pH value, sugar content, and oxygen concentration are crucial to the quality of fruit wine, and the traditional process is difficult to meet these fine and real-time control requirements.

[0007] In recent years, the application of Internet of Things and artificial intelligence technology in industrial production has gradually become popular, and many intelligent fermentation systems have emerged. These systems usually integrate sensor networks, data collection and analysis modules, and combine automatic control technology to realize real-time monitoring and control of the fermentation process. However, these systems are mostly still limited to fixed fermentation plants and are difficult to realize mobile deployment of the plant. At the same time, existing intelligent fermentation systems mostly focus on the application of a single algorithm, such as using machine learning models to predict the fermentation process and using deep learning to optimize fermentation parameters, but lack a global optimization strategy for the entire fermentation process, and cannot realize the overall intelligence and automation of the fermentation process.

[0008] Therefore, how to provide a deployment method for a mobile fruit wine fermentation plant based on modular integration is a problem that those skilled in the art need to solve. SUMMARY

[0009] One object of the present application is to provide a deployment method for a mobile fruit wine fermentation plant based on modular integration. The present application uses modular design, Internet of Things and advanced artificial intelligence algorithms to divide the fruit wine fermentation plant into independent fermentation modules, realizing intelligent and automated control of the entire process from fermentation start, main fermentation to aging. Through real-time monitoring and multi-layer coordination algorithms, the fermentation parameters are dynamically optimized to ensure accurate regulation of the fermentation process and high quality of the product. The plant can be flexibly deployed in different production areas, improving raw material utilization, reducing transportation costs, and continuously improving the fermentation process through data-driven methods, with the significant advantages of flexible deployment, efficient production, stable quality and automatic optimization.

[0010] The deployment method for a mobile fruit wine fermentation plant based on modular integration according to the embodiment of the present application comprises the following steps:

[0011] S1, dividing the plant into a first fermentation module, a second fermentation module and a third fermentation module, each fermentation module having a standardized interface and connection mode;

[0012] S2, introducing the fruit juice raw material into the first fermentation module, monitoring the sugar content, temperature and oxygen concentration using random forest and support vector machine algorithms, predicting the fermentation start conditions, adjusting the temperature and oxygen supply in the first fermentation module, and outputting the first fermentation liquid;

[0013] S3, introducing the first fermentation liquid into the second fermentation module, constructing a dynamic model of the fermentation process using graph convolution network and generative adversarial network, adjusting the temperature, stirring speed and oxygen supply of the second fermentation module, optimizing the conversion of sugar to alcohol, and outputting the second fermentation liquid;

[0014] S4, introducing the second fermentation liquid into the third fermentation module, analyzing the flavor substances using spatio-temporal graph neural network combined with variational autoencoder, predicting the optimal aging conditions, adjusting the micro-oxygen environment and temperature, completing the flavor optimization and clarification, and obtaining the finished fruit wine;

[0015] S5, installing a sensor network in each fermentation module to monitor temperature, pH, sugar content, alcohol concentration and oxygen concentration, constructing a global data set and transmitting it to the central control system;

[0016] S6, the central control system integrates a multi-layer coordination algorithm to coordinate the fermentation parameters between the fermentation modules using the received global data set to optimize the overall control of the fermentation process;

[0017] S7, after fermentation is completed, an automatic cleaning and disinfection system is used to clean and disinfect each fermentation module;

[0018] S8, when mobile deployment is needed, disassemble the fermentation modules, load them onto a mobile carrier, move to the target location through a transportation tool, and complete the reassembly and connection of the fermentation modules.

[0019] Optionally, the S2 specifically includes:

[0020] S21, installing sensors in the first fermentation module to collect data on sugar content, temperature and oxygen concentration;

[0021] S22, using random forest algorithm to select features from historical fermentation data to determine key factors affecting fermentation start;

[0022] S23, based on the selected key factors, constructing a support vector machine prediction model to predict the fermentation start conditions:

[0023]

[0024] where g(x) represents the prediction function, sign represents the sign function, N represents the total number of input vectors, α i represents the weight of the support vector, y i represents the classification label corresponding to the support vector, and K(xi x) represents a kernel function, x i represents a support vector in the training data, b represents a bias term, and x represents a currently measured input vector;

[0025] y i ∈{+1,-1}, where +1 represents that the fermentation start condition is met, and -1 represents that the fermentation start condition is not met;

[0026] x is composed of sugar content, temperature, and oxygen concentration, x=[x 糖度 ,x 温度 ,x 氧气浓度 ];

[0027] S24, input the current real-time measured input vector x into the support vector machine prediction model g(x) to obtain a prediction result:

[0028] If g(x)=+1, it indicates that the current condition meets the fermentation start requirement;

[0029] If g(x)=-1, it indicates that the current condition does not meet the fermentation start requirement and needs to be adjusted;

[0030] S25, according to the prediction result, dynamically adjust the temperature and oxygen supply in the first fermentation module by using a control system:

[0031] When g(x)=-1, calculate the parameter increment Δx 温度 and Δx 氧气浓度 ;

[0032] Adjust the temperature to x 温度 +Δx 温度 by a temperature control device;

[0033] Adjust the oxygen concentration to x 氧气浓度 +Δx 氧气浓度 by an oxygen supply system;

[0034] S26, repeat steps S21 to S25 until the support vector machine prediction model outputs g(x)=+1, and the generation of the first fermentation broth is completed.

[0035] Optionally, the S3 specifically comprises:

[0036] S31, introduce the first fermentation broth into a second fermentation module, dynamically model the parameters in the fermentation broth by using a graph convolution network, and construct a multi-dimensional relationship graph G=(V,E) of the fermentation process, wherein the node set V represents key fermentation parameters, including temperature, sugar content, alcohol concentration, pH value, and oxygen concentration, the edge set E represents the mutual influence relationship between the parameters, and each node v i ∈V has an initial feature vector h i, update the feature matrix of each node through multi-layer propagation of the graph convolution network:

[0037]

[0038] where H (l+1) represents the feature matrix of the l+1th layer, H (l) represents the feature matrix of the lth layer, and σ represents an activation function, represents the degree matrix of , represents the graph adjacency matrix with a self-loop, W (l) represents the trainable weight matrix of the lth layer, and α k represents the weight coefficient, R k represents the matrix of the kth relationship type, and K represents the total number of relationship type matrices.

[0039] S32, simulate the nonlinear dynamic change of sugar to alcohol conversion in the fermentation process by using a generative adversarial network, the generative adversarial network comprising a generator G and a discriminator D, wherein the generator G is composed of a multi-layer adaptive normalization and a residual network, and is used to generate a fermentation parameter sequence G(z):

[0040] G(z) = ResNet(AdaIN(z, S style ));

[0041] where z represents an input random noise vector, S style represents a fermentation condition style vector extracted by a style encoder, AdaIN represents adaptive normalization, and ResNet represents a residual network.

[0042] S33, the discriminator D adopts a multi-scale convolution network to discriminate the actual fermentation parameter sequence x and the generated fermentation parameter sequence G(z) from different scales, and optimize the objective function.

[0043] S34, after training, the generator G generates an optimal fermentation parameter sequence, and combines the multi-dimensional feature correlation output by the graph convolution network to obtain an optimal parameter adjustment vector ΔP = [ΔT, ΔS, ΔO] at each time step:

[0044] ΔP t = G(z) + GCN(H (l) );

[0045] where ΔP t represents the optimal parameter adjustment vector at the tth time step, ΔT represents the temperature adjustment value, ΔS represents the stirring speed adjustment value, ΔO represents the oxygen concentration adjustment value, and GCN represents the graph convolution network.

[0046] S35, according to the optimal parameter adjustment vector, the control system of the second fermentation module is used to adjust in real time:

[0047] Adjust the temperature to T t + ΔT;

[0048] Adjust the stirring speed to S t + ΔS;

[0049] Adjust the oxygen concentration to O t + ΔO;

[0050] S36, repeatedly performing steps S31 to S35 to continuously optimize the fermentation process of the second fermentation module until the second fermentation liquor is output.

[0051] Optionally, the S4 specifically comprises:

[0052] S41, introducing the second fermentation liquor into a third fermentation module, modeling the spatiotemporal dynamic characteristics of the flavor substances in the fermentation liquor using a spatiotemporal graph neural network, and constructing a multi-dimensional spatiotemporal feature map G s = (V s , E s ) of the fermentation liquor, where V s represents a set of key flavor substance parameter nodes at time s, including acidity, ester and phenolic substance concentration, E s represents the spatiotemporal association relationship between parameters, and the node feature matrix is updated through multi-layer propagation of the spatiotemporal graph neural network

[0053]

[0054] wherein, represents the feature matrix of the q+1 layer at time s, σ represents an activation function, N(j) represents a set of neighbor nodes of node j, η jk represents the spatiotemporal weight between node j and node k, P (q) represents a trainable weight matrix of the spatial convolution kernel, represents a trainable weight matrix of the temporal convolution kernel, Q r represents a feature transfer matrix, δ r represents the feature transfer matrix Q r the influence weight of the flavor substance, represents the feature matrix of the q layer at time s, and R represents the number of feature transfer matrices;

[0055] S42, dimensionality reduction and reconstruction of the feature space of the flavor substances using a variational autoencoder, the variational autoencoder comprising an encoder and a decoder, the encoder mapping the input high-dimensional feature vector Z s into a latent space U, and the decoder reconstructing the latent variable u into the original feature space, the latent space distribution of the encoder being:

[0056] p ψ (u|Z s )=N(u|μ(Z s ),Σ(Z s ));

[0057] wherein, p ψ (u|Z s ) represents the probability distribution of latent variable u given input Z s , μ(Z s ) represents the mean of latent variable u, Σ(Z s ) represents the covariance matrix of latent variable u, and N(u|μ(Z s ),Σ(Z s )) represents a normal distribution;

[0058] S43, sampling the latent variable by variational inference to obtain the optimal feature expression of flavor substances, combining the spatio-temporal graph neural network to model the spatio-temporal dynamics of parameters, obtaining the optimal aging condition, including the adjustment parameters of the micro-oxygen environment and temperature, constructing the optimal parameter adjustment vector ΔV s =[ΔO,ΔR], wherein:

[0059] ΔO represents the micro-oxygen environment adjustment amount, used to adjust the oxygen concentration in the fermentation broth;

[0060] ΔR represents the temperature adjustment amount, used to adjust the temperature in the fermentation broth;

[0061] S44, according to the optimal parameter adjustment vector, using the control system of the third fermentation module to perform real-time adjustment:

[0062] adjusting the micro-oxygen environment to O s +ΔO, wherein O s represents the oxygen concentration at time s;

[0063] adjusting the temperature to R s +ΔR, wherein R s represents the temperature at time s;

[0064] S45, repeatedly executing steps S41 to S44 to continuously optimize the aging process of the third fermentation module until the flavor optimization and clarification treatment are completed, and the finished fruit wine is obtained.

[0065] Optionally, the S6 specifically comprises:

[0066] S61, transmitting the real-time data collected by the sensor network in each fermentation module to the central control system to construct a global data set D g{X1,X2,X3}, wherein X1 represents real-time monitoring data of the first fermentation module, X2 represents real-time monitoring data of the second fermentation module, and X3 represents real-time monitoring data of the third fermentation module;

[0067] S62, using a multi-layer coordination algorithm to process the global data set D g , a method combining reinforcement learning and Bayesian optimization is used to define the state space S t , action space A t and reward function R t ;

[0068] S63, in the reinforcement learning framework, a deep Q network is used to learn and optimize the parameter adjustment strategy, the goal is to maximize the cumulative reward function R t :

[0069]

[0070] wherein Q * (S t ,A t ) represents the optimal Q value of performing action A t under state S t , π represents the probability of selecting an action at each state, and γ represents the discount factor;

[0071] S64, using Bayesian optimization to further fine-tune the parameters θ of each fermentation module, using a Gaussian process model to approximate the objective function, and selecting the next evaluation point by maximizing the expected improvement value:

[0072]

[0073] wherein θ t+1 represents the next parameter configuration, and EI(θ) represents the expected improvement value;

[0074] S65, feeding back the optimal parameter adjustment scheme obtained through reinforcement learning and Bayesian optimization to each fermentation module to adjust the fermentation conditions of each module in real time;

[0075] S66, repeating steps S61 to S65 to continuously coordinate and optimize the entire fermentation process.

[0076] Optionally, the S62 specifically includes:

[0077] S621, the state space S t is composed of the current parameter state of each fermentation module, and is represented as S t ={s1,s2,s3}, wherein s i represents the parameter state of the i-th fermentation module.

[0078] S622, Action space A t The parameter adjustment operation executable by the central control system is denoted as A t = {a1, a2, a3}, where a i represents the parameter adjustment operation performed on the i-th fermentation module;

[0079] S623, Reward function R t is defined as a comprehensive evaluation function of the sugar conversion efficiency, alcohol generation rate and flavor formation degree in the fermentation process:

[0080] R t = w1E(sugar conversion) + w2E(alcohol generation) + w3E(flavor balance);

[0081] where w1, w2 and w3 represent weight coefficients, E(sugar conversion) represents the efficiency of sugar conversion to alcohol, E(alcohol generation) represents the alcohol generation rate, and E(flavor balance) represents the balance degree of flavor substances;

[0082]

[0083] where, represents the instantaneous alcohol generation rate, represents the instantaneous by-product generation rate, C 糖 (t) represents the sugar concentration at time t, represents the acidity change rate, ρ1 and ρ2 represent weight coefficients, τ represents a regulation coefficient, ∈ represents a smoothing term, t1 represents the termination time of integration, and t0 represents the starting time of integration;

[0084]

[0085] where M represents the number of observation samples, θ0 represents the constant term of the model, ΔC 酒精 (t-p) represents the change in alcohol concentration at time t-p, θ p represents the autoregressive coefficient of the autoregressive moving average model, φ q represents the moving average coefficient of the autoregressive moving average model, ε t-q represents the random disturbance term at time t-p, P represents the order of the autoregressive model part, Q represents the order of the moving average model part, δ represents the adjustment strength coefficient of the influence function f(T, O2, S) on the alcohol generation rate, T represents the temperature, O2 represents the oxygen concentration, and S represents the stirring rate:

[0086]

[0087] where α T , α O2 and αS Indicates the influence coefficient;

[0088]

[0089] Where exp represents the exponential function, H(C 风味 ξ represents the complexity of flavor compounds. i I(C) represents the weighting coefficient. 风味,i C 目标风味,i ) represents the mutual information between the current concentration of the i-th flavor compound and the target flavor concentration;

[0090]

[0091] Wherein, p(C 风味,i ) represents the probability distribution of the concentration of the i-th flavor compound;

[0092]

[0093] Wherein, p(c i ,c j Let p(c) represent the joint probability distribution of the i-th flavor compound and the j-th target flavor. i Let p(c) represent the marginal probability distribution of the i-th flavor compound. j Let represent the marginal probability distribution of the j-th target flavor.

[0094] The beneficial effects of this invention are:

[0095] In terms of fermentation process control, this invention employs advanced artificial intelligence algorithms to achieve intelligent, automated, and precise control of the entire fermentation process. Through random forest and support vector machine algorithms, the factory can monitor key parameters in real time, accurately predict fermentation start-up conditions, and optimize the initial stage of fermentation. Graph convolutional networks and generative adversarial networks are used to dynamically model the fermentation process, enabling precise control of the sugar-to-alcohol conversion during the mid-fermentation stage to ensure a balance between alcohol production rate and byproducts. In the later stages of fermentation, spatiotemporal graph neural networks and variational autoencoders analyze the dynamic characteristics of flavor compounds, predicting and adjusting optimal aging conditions to ensure the complexity and uniqueness of the fruit wine's flavor. The entire fermentation process is globally optimized through a multi-layered coordination algorithm in the central control system, ensuring coordinated parameter cooperation between fermentation modules, achieving real-time control and overall optimization of the fermentation process, and guaranteeing high product quality and consistency.

[0096] In terms of data collection and analysis, the application installs a sensor network in each fermentation module to monitor key parameters such as temperature, pH value, sugar content, alcohol concentration, and oxygen concentration in real time during the fermentation process, forming a complete global data set. Through real-time analysis of these data by the central control system, combined with algorithms such as reinforcement learning and Bayesian optimization, the process parameters of each fermentation stage are continuously optimized to ensure that the fermentation process runs in the optimal state. At the same time, using big data analysis and machine learning technology, the application can continuously learn and improve the fermentation strategy from historical fermentation data, realizing data-driven fermentation process improvement and enabling the factory to have the ability of self-learning and continuous optimization.

[0097] Through the automatic cleaning and disinfection system, the application also effectively solves the problem of high difficulty in cleaning and disinfection and high manual intervention in traditional processes. After fermentation is completed, each fermentation module can be automatically cleaned and disinfected, avoiding the risk of cross contamination and improving the hygiene safety of the product. In addition, the modular design enables the factory to be quickly disassembled and reorganized when mobile deployment is needed, flexibly responding to changes in production location requirements, further reducing production costs and improving the operating efficiency of the factory. BRIEF DESCRIPTION OF DRAWINGS

[0098] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0099] Figure 1 Flow chart of the deployment method of the mobile fruit wine fermentation factory based on modular integration proposed by the application;

[0100] Figure 2 Deployment structure schematic diagram of the deployment method of the mobile fruit wine fermentation factory based on modular integration proposed by the application. DETAILED DESCRIPTION

[0101] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams that only schematically illustrate the basic structure of the application, and therefore only show the components related to the application.

[0102] Reference Figure 1 and Figure 2 The deployment method of the mobile fruit wine fermentation factory based on modular integration includes the following steps:

[0103] S1, the factory is divided into a first fermentation module, a second fermentation module, and a third fermentation module, each fermentation module having a standardized interface and connection method;

[0104] S2, introducing the fruit juice raw material into the first fermentation module, monitoring the sugar content, temperature and oxygen concentration using random forest and support vector machine algorithms, predicting the fermentation start conditions, adjusting the temperature and oxygen supply in the first fermentation module, and outputting the first fermentation liquor;

[0105] S3, introducing the first fermentation liquor into the second fermentation module, constructing a dynamic model of the fermentation process using graph convolution network and generative adversarial network, adjusting the temperature, stirring speed and oxygen supply of the second fermentation module, optimizing the conversion of sugar to alcohol, and outputting the second fermentation liquor;

[0106] S4, introducing the second fermentation liquor into the third fermentation module, analyzing the flavor substances using spatio-temporal graph neural network combined with variational autoencoder, predicting the optimal aging conditions, adjusting the micro-oxygen environment and temperature, completing the flavor optimization and clarification, and obtaining the finished fruit wine;

[0107] S5, installing a sensor network in each fermentation module to monitor temperature, pH, sugar content, alcohol concentration and oxygen concentration, constructing a global data set and transmitting it to the central control system;

[0108] S6, the central control system integrates a multi-layer coordination algorithm to coordinate the fermentation parameters between the fermentation modules using the received global data set to optimize the overall control of the fermentation process;

[0109] S7, after fermentation is completed, an automatic cleaning and disinfection system is used to clean and disinfect each fermentation module;

[0110] S8, when mobile deployment is needed, disassemble the fermentation modules, load them onto a mobile carrier, move to the target location by transportation means, and complete the reassembly and connection of the fermentation modules.

[0111] In this embodiment, S2 specifically includes:

[0112] S21, installing sensors in the first fermentation module to collect data on sugar content, temperature and oxygen concentration;

[0113] S22, using random forest algorithm to select features from historical fermentation data to determine key factors affecting fermentation start;

[0114] S23, based on the selected key factors, constructing a support vector machine prediction model to predict the fermentation start conditions:

[0115]

[0116] where g(x) represents the prediction function, sign represents the sign function, N represents the total number of input vectors, α i represents the weight of the support vector, y i represents the classification label corresponding to the support vector, K(xi x) represents a kernel function, x i represents a support vector in the training data, b represents a bias term, and x represents a currently measured input vector;

[0117] y i ∈{+1,-1}, where +1 represents that the fermentation start condition is met, and -1 represents that the fermentation start condition is not met;

[0118] x is composed of sugar content, temperature, and oxygen concentration, x = [x 糖度 ,x 温度 ,x 氧气浓度 ];

[0119] S24, input the current real-time measured input vector x into the support vector machine prediction model g(x) to obtain a prediction result:

[0120] If g(x) = +1, it indicates that the current condition meets the fermentation start requirement;

[0121] If g(x) = -1, it indicates that the current condition does not meet the fermentation start requirement and needs to be adjusted;

[0122] S25, according to the prediction result, dynamically adjust the temperature and oxygen supply in the first fermentation module by using a control system:

[0123] When g(x) = -1, calculate the parameter increment Δx 温度 and Δx 氧气浓度 that need to be adjusted;

[0124] Adjust the temperature to x 温度 + Δx 温度 by a temperature control device;

[0125] Adjust the oxygen concentration to x 氧气浓度 + Δx 氧气浓度 by an oxygen supply system;

[0126] S26, repeat steps S21 to S25 until the support vector machine prediction model outputs g(x) = +1, and the generation of the first fermentation broth is completed.

[0127] In the embodiment, S3 specifically includes:

[0128] S31, introduce the first fermentation broth into the second fermentation module, dynamically model the parameters in the fermentation broth by using a graph convolution network, and construct a multi-dimensional relationship graph G = (V, E) of the fermentation process, where the node set V represents key fermentation parameters, including temperature, sugar content, alcohol concentration, pH value, and oxygen concentration, the edge set E represents the mutual influence relationship between the parameters, and each node v i ∈ V has an initial feature vector h i, update the feature matrix of each node through multi-layer propagation of the graph convolution network:

[0129]

[0130] where H (l+1) represents the feature matrix of the l+1th layer, H (l) represents the feature matrix of the lth layer, and σ represents an activation function, represents the degree matrix of , represents the graph adjacency matrix with a self-loop, W (l) represents the trainable weight matrix of the lth layer, and α k represents the weight coefficient, R k represents the matrix of the kth relationship type, and K represents the total number of relationship type matrices.

[0131] S32, simulate the nonlinear dynamic change of sugar to alcohol conversion in the fermentation process by using a generative adversarial network, the generative adversarial network comprising a generator G and a discriminator D, wherein the generator G is composed of a multi-layer adaptive normalization and a residual network, and is used to generate a fermentation parameter sequence G(z):

[0132] G(z) = ResNet(AdaIN(z, S style ));

[0133] where z represents an input random noise vector, S style represents a fermentation condition style vector extracted by a style encoder, AdaIN represents adaptive normalization, and ResNet represents a residual network.

[0134] S33, the discriminator D adopts a multi-scale convolution network to discriminate the actual fermentation parameter sequence x and the generated fermentation parameter sequence G(z) from different scales, and optimize the objective function.

[0135] S34, after training, the generator G generates an optimal fermentation parameter sequence, and combines the multi-dimensional feature correlation output by the graph convolution network to obtain an optimal parameter adjustment vector ΔP = [ΔT, ΔS, ΔO] at each time step:

[0136] ΔP t = G(z) + GCN(H (l) );

[0137] where ΔP t represents the optimal parameter adjustment vector at the tth time step, ΔT represents the temperature adjustment value, ΔS represents the stirring speed adjustment value, ΔO represents the oxygen concentration adjustment value, and GCN represents the graph convolution network.

[0138] S35, according to the optimal parameter adjustment vector, the control system of the second fermentation module is used to adjust in real time:

[0139] Adjust the temperature to T t + ΔT;

[0140] Adjust the stirring speed to S t + ΔS;

[0141] Adjust the oxygen concentration to O t + ΔO;

[0142] S36, repeatedly performing steps S31 to S35 to continuously optimize the fermentation process of the second fermentation module until the second fermentation liquor is output.

[0143] In this embodiment, S4 specifically comprises:

[0144] S41, introducing the second fermentation liquor into a third fermentation module, modeling the spatiotemporal dynamic characteristics of flavor substances in the fermentation liquor using a spatiotemporal graph neural network, and constructing a multi-dimensional spatiotemporal feature map G s of the fermentation liquor s , s wherein V s represents a set of key flavor substance parameter nodes at time s, including acidity, ester and phenolic substance concentrations, E s represents the spatiotemporal association relationship between parameters, and the node feature matrix is updated through multi-layer propagation of the spatiotemporal graph neural network

[0145]

[0146] wherein, represents the feature matrix of the q+1th layer at time s, σ represents an activation function, N(j) represents a set of neighbor nodes of node j, η jk represents the spatiotemporal weight between node j and node k, P (q) represents a trainable weight matrix of the spatial convolution kernel, represents a trainable weight matrix of the temporal convolution kernel, Q r represents a feature transfer matrix, δ r represents the feature transfer matrix Q r of the influence weight of the flavor substance, represents the feature matrix of the qth layer at time s, and R represents the number of feature transfer matrices;

[0147] S42, dimensionality reduction and reconstruction of the feature space of the flavor substance using a variational autoencoder, the variational autoencoder comprising an encoder and a decoder, the encoder mapping the input high-dimensional feature vector Z s into a latent space U, and the decoder reconstructing the latent variable u into the original feature space, the latent space distribution of the encoder being:

[0148] p ψ (u|Z s )=N(u|μ(Z s ),Σ(Z s ));

[0149] Where, p ψ (u|Z s ) represents the given input Z s The probability distribution of the latent variable u, μ(Z) s ) represents the mean of the latent variable u, Σ(Z) s N(u|μ(Z)) represents the covariance matrix of the latent variable u. s ),Σ(Z s )) represents a normal distribution;

[0150] S43. By sampling latent variables through variational inference, the optimal characteristic expression of flavor substances is obtained. Combined with spatiotemporal graph neural network, the spatiotemporal dynamics of parameters are modeled to obtain the optimal aging conditions, including the adjustment parameters of micro-oxygen environment and temperature, and the optimal parameter adjustment vector ΔV is constructed. s = [ΔO, ΔR], where:

[0151] ΔO represents the microaerobic environment adjustment amount, which is used to adjust the oxygen concentration in the fermentation broth;

[0152] ΔR represents the temperature adjustment amount, used to regulate the temperature in the fermentation broth;

[0153] S44. Adjust the vector according to the optimal parameters and make real-time adjustments using the control system of the third fermentation module:

[0154] Adjust the micro-oxygen environment to O s +ΔO, where O s This represents the oxygen concentration at time s;

[0155] Adjust the temperature to R s +ΔR, where R s This represents the temperature at time s.

[0156] S45. Repeat steps S41 to S44 to continuously optimize the aging process of the third fermentation module until flavor optimization and clarification are completed to obtain the finished fruit wine.

[0157] In this embodiment, S6 specifically includes:

[0158] S61. Transmit the real-time data collected by the sensor network in each fermentation module to the central control system to construct a global dataset D. g{X1,X2,X3}, wherein X1 represents real-time monitoring data of the first fermentation module, X2 represents real-time monitoring data of the second fermentation module, and X3 represents real-time monitoring data of the third fermentation module;

[0159] S62, using a multi-layer coordination algorithm to process the global data set D g , a method combining reinforcement learning and Bayesian optimization is used to define the state space S t , action space A t and reward function R t ;

[0160] S63, in the reinforcement learning framework, using a deep Q network to learn and optimize the parameter adjustment strategy, the goal is to maximize the cumulative reward function R t :

[0161]

[0162] wherein Q * (S t ,A t ) represents the optimal Q value of performing action A t in state S t , π represents the probability of selecting an action at each state, and γ represents the discount factor;

[0163] S64, using Bayesian optimization to further fine-tune the parameters θ of each fermentation module, using a Gaussian process model to approximate the objective function, and selecting the next evaluation point by maximizing the expected improvement value:

[0164]

[0165] wherein θ t+1 represents the next parameter configuration, and EI(θ) represents the expected improvement value;

[0166] S65, feeding back the optimal parameter adjustment scheme obtained through reinforcement learning and Bayesian optimization to each fermentation module to adjust the fermentation conditions of each module in real time;

[0167] S66, repeating steps S61 to S65 to continuously coordinate and optimize the entire fermentation process.

[0168] In the present embodiment, S62 specifically includes:

[0169] S621, the state space S t is composed of the current parameter state of each fermentation module, represented as S t ={s1,s2,s3}, wherein s i represents the parameter state of the i-th fermentation module;

[0170] S622, Action space A t The parameter adjustment operations executable by the central control system constitute an action space A t = {a1, a2, a3}, where a i represents the parameter adjustment operation performed on the i-th fermentation module;

[0171] S623, Reward function R t is defined as a comprehensive evaluation function of the sugar conversion efficiency, alcohol generation rate and flavor formation degree in the fermentation process:

[0172] R t = w1E(sugar conversion) + w2E(alcohol generation) + w3E(flavor balance);

[0173] where w1, w2 and w3 represent weight coefficients, E(sugar conversion) represents the efficiency of sugar conversion to alcohol, E(alcohol generation) represents the alcohol generation rate, and E(flavor balance) represents the balance degree of flavor substances;

[0174]

[0175] where, represents the instantaneous alcohol generation rate, represents the instantaneous by-product generation rate, C 糖 (t) represents the sugar concentration at time t, represents the rate of change of acidity, ρ1 and ρ2 represent weight coefficients, τ represents a regulation coefficient, ∈ represents a smoothing term, t1 represents the termination time of integration, and t0 represents the starting time of integration;

[0176]

[0177] where M represents the number of observation samples, θ0 represents the constant term of the model, ΔC 酒精 (t-p) represents the change in alcohol concentration at time t-p, θ p represents the autoregressive coefficient of the autoregressive moving average model, φ q represents the moving average coefficient of the autoregressive moving average model, ε t-q represents the random disturbance term at time t-p, P represents the order of the autoregressive model part, Q represents the order of the moving average model part, δ represents the adjustment strength coefficient of the influence function f(T, O2, S) on the alcohol generation rate, T represents the temperature, O2 represents the oxygen concentration, and S represents the stirring rate:

[0178]

[0179] where α T , and αS represents an impact coefficient;

[0180]

[0181] wherein exp represents an exponential function, H(C 风味 ) represents the complexity of the flavor substance, ξ i represents a weight coefficient, I(C 风味,i ; C 目标风味,i ) represents the mutual information between the current concentration of the i-th flavor substance and the target flavor concentration;

[0182]

[0183] wherein p(C 风味,i ) represents the distribution probability of the i-th flavor substance concentration;

[0184]

[0185] wherein p(c i ,c j ) represents the joint probability distribution of the i-th flavor substance and the j-th target flavor, p(c i ) represents the marginal probability distribution of the i-th flavor substance, and p(c j ) represents the marginal probability distribution of the j-th target flavor.

[0186] Example 1:

[0187] In order to verify the feasibility of the application in implementation, the application is applied to the vicinity of a grape vineyard in a certain fruit wine producing area. In order to improve the efficiency of fruit wine production and reduce the transportation cost of raw materials, it is decided to deploy a mobile fruit wine fermentation plant based on modular integration. Traditionally, fruit wine production in this area relies on fixed fermentation plants, and after grape picking, long-distance transportation to fixed plants for fermentation is required. This not only increases transportation costs, but also leads to a decrease in the freshness of raw materials, thereby affecting the quality of fruit wine. In order to solve these problems, the modular mobile fruit wine fermentation plant of the application is used, which is quickly deployed near the vineyard and fully utilizes artificial intelligence and automatic control technology to realize the intelligentization and high efficiency of fruit wine fermentation.

[0188] During the grape harvesting season, to make the most of the freshness of the grapes, the mobile fermentation plant is transported by truck and quickly deployed on an empty lot next to the vineyard. The plant consists of three main fermentation modules: the first fermentation module, the second fermentation module, and the third fermentation module. First, fresh grapes are crushed, and the resulting juice is directly introduced into the first fermentation module. This module monitors the sugar content, temperature, and oxygen concentration in the juice in real time through the installed sensors. Using random forest and support vector machine algorithms, these data are analyzed to predict the optimal conditions for the start of fermentation. In this scenario, the fermentation start time is shortened by an average of 12 hours, from the traditional 48 hours of initial fermentation to about 36 hours. The automated control system adjusts the temperature and oxygen supply based on the prediction results to ensure that the fermentation start is under optimal conditions, thereby improving fermentation efficiency and reducing human intervention.

[0189] When the first fermentation liquid is generated, it is automatically transported to the second fermentation module. In this stage, the plant uses graph convolution networks and generative adversarial networks to model and optimize the fermentation process. Traditional fermentation processes have difficulty accurately controlling the rate of sugar conversion to alcohol, often resulting in fermentation times of 5 to 7 days. However, this plant uses deep learning models to analyze parameters such as sugar content, alcohol concentration, and pH in the fermentation liquid in real time, enabling dynamic optimization of fermentation parameters. During the experiment, the plant successfully completed the conversion of sugar in 72 hours, with the alcohol concentration in the fermentation liquid reaching the expected 12%, while the generation of byproducts such as acetic acid was controlled below 0.05%, far lower than the 0.1% to 0.2% in traditional processes. By adjusting the temperature, stirring speed, and oxygen supply in real time, this module achieves fine control over the sugar conversion process, not only shortening the fermentation time but also effectively increasing alcohol yield and flavor stability.

[0190] The second fermentation liquid then enters the third fermentation module, where the plant uses spatio-temporal graph neural networks combined with variational autoencoders to dynamically analyze and optimize flavor substances in the fermentation liquid. During fermentation, the formation of flavor substances such as esters and phenols is crucial to the taste and quality of the wine. In traditional processes, the aging process usually takes 1 to 3 months, and flavor formation is highly uncertain and easily affected by temperature and oxygen changes. The plant uses neural network models to predict the optimal aging conditions and uses an automated system to adjust the micro-oxygen environment and temperature in real time. Experimental data show that after 21 days of optimized aging, the ester content in the wine reaches 150 mg / L, and the phenol content stabilizes at 250 mg / L, both of which are superior to the 100 mg / L and 200 mg / L in traditional processes. By optimizing flavor balance, the produced wine scores an average of 15% higher in taste tests, with significantly enhanced flavor complexity and aroma intensity.

[0191] The entire fermentation process is globally monitored and coordinated by a central control system. A sensor network continuously collects data such as temperature, pH, sugar content, alcohol concentration, and oxygen concentration in each module and transmits them to the central control system. The system integrates multi-layer coordination algorithms and uses reinforcement learning and Bayesian optimization to globally optimize the parameters of each fermentation module. In the experiment, the central control system adjusts the fermentation parameters in real time, increasing the fermentation efficiency of the entire factory by 20% and reducing energy consumption by 15%. Through big data analysis, the factory generates 5000 liters of wine in one fermentation cycle (about 30 days), reducing the production cost of each liter of wine by 30%.

[0192] After fermentation is complete, an automated cleaning and disinfection system automatically cleans and disinfects each fermentation module. This process does not require human intervention and takes only 2 hours, reducing the time by 50% compared to traditional processes. The fast cleaning process ensures the hygiene of the production environment and avoids the risk of cross-contamination, preparing for the next batch of fermentation. In addition, the factory completes the disassembly and loading of the modules in only 1 day after the grape picking season ends, moves them to the next vineyard for deployment, and resumes fermentation production. This flexible deployment method not only reduces transportation costs but also ensures the freshness of raw materials and the quality of wine.

[0193] Table 1 Comparison of modular integrated mobile wine fermentation plant and traditional fermentation process

[0194] Indicators Traditional fruit wine fermentation process Mobile fruit wine fermentation plant (the invention) Fermentation start-up time 48 hours 36 hours Primary fermentation time 5-7 days 72 hours Alcohol concentration 10-12％ 12％ By-product generation (acetic acid) 0.1％-0.2％ 0.05％ Aging time 1-3 months 21 days Ester content 100 mg / L 150 mg / L Phenolic content 200 mg / L 250 mg / L Flavor complexity improvement - 15% increase Fermentation efficiency improvement - 20% increase Energy consumption - 15% reduction Cost per liter of fruit wine production - 30% reduction Automated cleaning cycle 4 hours 2 hours Cleaning time reduction - 50％ Deployment time Fixed plant, no need to move 1 day Raw material transportation cost High Low Fruit wine production during fermentation cycle 4000 liters 5000 liters Manual intervention High Low, automated control Production environment hygiene Manual cleaning, high risk Automatic cleaning, no cross-contamination risk

[0195] In the above Table 1, the advantages of the modular integrated mobile wine fermentation plant compared to the traditional wine fermentation process are clearly demonstrated, especially in terms of fermentation time, efficiency, product quality, and other key aspects.

[0196] First, in terms of fermentation start-up time, the traditional process usually takes 48 hours, while the mobile fermentation plant only takes 36 hours, significantly shortening the fermentation start-up time. This is mainly due to the application of automated monitoring and control systems, which monitor fermentation conditions in real time and make optimal adjustments, greatly improving initial fermentation efficiency.

[0197] In terms of initial fermentation time, the traditional process usually takes 5 to 7 days to complete, while the mobile plant can complete fermentation in 72 hours with the support of intelligent optimization algorithms and deep learning technology, greatly shortening fermentation time. In terms of alcohol production, the alcohol concentration in the traditional process is between 10%-12%, while the mobile plant can accurately control it to 12%, and the amount of byproduct generation (such as acetic acid) is also significantly reduced, only 0.05%, more than half of the traditional process of 0.1%-0.2%.

[0198] In terms of aging time, the traditional process usually takes 1 to 3 months, while the mobile fermentation plant can complete the aging process in 21 days through precise control of the micro-oxygen environment and temperature. At the same time, the ester content of the fruit wine produced by the mobile plant reaches 150 mg / L, and the phenolic substance content is stable at 250 mg / L, which is better than 100 mg / L and 200 mg / L under the traditional process, further proving the significant effect of the invention on flavor optimization. In addition, the flavor complexity is improved by 15%, and through the optimization of the process and intelligent control, the flavor of the fruit wine is more rich and stable.

[0199] In terms of fermentation efficiency, the efficiency of the mobile plant is improved by 20%, and the energy consumption is reduced by 15%, which means that more fruit wine can be produced with less energy in the same time. The production cost per liter of fruit wine is reduced by 30%, which greatly reduces the production cost.

[0200] In addition, in the cleaning and disinfection process, the mobile plant uses an automatic cleaning system to shorten the cleaning cycle to 2 hours, which is 50% less than the 4 hours of the traditional process, effectively improving the production efficiency. At the same time, the plant deployment time is only 1 day, which can quickly respond to market demand, while the traditional fixed plant cannot be moved and has poor flexibility.

[0201] In the production process, the traditional process relies on manual monitoring and operation, which has high labor cost and error, while the mobile plant uses an intelligent control system to realize low human intervention, ensuring the stability of the production process. In addition, the automatic cleaning system avoids the risk of cross contamination, significantly improving the hygiene standard of the production environment.

[0202] Overall, the mobile fruit wine fermentation plant is superior to the traditional fruit wine fermentation process in terms of fermentation time, product quality, production cost and production efficiency, etc., and shows its significant advantages in intelligence, flexibility and efficiency. These comparison data show that the mobile plant can provide higher economic benefits and product quality for fruit wine production, especially when facing different production areas and market demands, its flexible deployment capability also brings wider development space for the fruit wine industry.

[0203] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this. Any skilled person in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. Method for the deployment of a mobile fruit wine fermentation plant based on modular integration, characterized in that, The method comprises the following steps: S1, dividing the factory into a first fermentation module, a second fermentation module and a third fermentation module, each fermentation module having a standardized interface and connection mode; S2, introducing the fruit juice raw material into the first fermentation module, monitoring the sugar content, temperature and oxygen concentration, predicting the fermentation start conditions, adjusting the temperature and oxygen supply in the first fermentation module, and outputting the first fermentation liquor; S3, introducing the first fermentation liquor into the second fermentation module, constructing a dynamic model of the fermentation process using a graph convolution network and a generative adversarial network, adjusting the temperature, stirring speed and oxygen supply of the second fermentation module, optimizing the conversion of sugar to alcohol, and outputting the second fermentation liquor; S4, introducing the second fermentation liquor into the third fermentation module, analyzing the flavor substances using a spatio-temporal graph neural network combined with a variational autoencoder, predicting the optimal aging conditions, adjusting the micro-oxygen environment and temperature, completing flavor optimization and clarification, and obtaining the finished fruit wine; S5, installing a sensor network in each fermentation module to monitor temperature, pH, sugar content, alcohol concentration and oxygen concentration, constructing a global data set and transmitting it to the central control system; S6, the central control system integrates a multi-layer coordination algorithm to coordinate the fermentation parameters between the fermentation modules using the received global data set to optimize the overall control of the fermentation process; S7, after fermentation is completed, an automatic cleaning and disinfection system is used to clean and disinfect each fermentation module; S8, when mobile deployment is required, disassemble the fermentation modules, load them onto a mobile carrier, move them to the target location by transportation means, and complete the reassembly and connection of the fermentation modules.

2. The method for deployment of a mobile fruit wine fermentation plant based on modular integration according to claim 1, characterized in that, The S2 specifically comprises: S21, installing sensors in the first fermentation module to collect data on sugar content, temperature and oxygen concentration; S22, using a random forest algorithm to select features from historical fermentation data to determine key factors affecting fermentation start; S23, based on the selected key factors, constructing a support vector machine prediction model to predict fermentation start conditions: where g(x) represents a prediction function, sign represents a sign function, N represents a total number of input vectors, a i represents a weight of a support vector, y i represents a classification label corresponding to the support vector, K(x i , x) represents a kernel function, x i represents a support vector in training data, b represents a bias term, and x represents a currently measured input vector; y i ∈ {+1, -1}, wherein +1 indicates that the fermentation start condition is satisfied, and -1 indicates that the fermentation start condition is not satisfied; x is composed of sugar content, temperature and oxygen concentration, x = [x 糖度 ,x 温度 ,x 氧气浓度 ] S24, inputting the current real-time measured input vector x into the support vector machine prediction model g(x) to obtain the prediction result: If g(x) = +1, it indicates that the current conditions meet the requirements for fermentation start; If g(x) = -1, it indicates that the current conditions do not meet the requirements for fermentation start and need to be adjusted; S25, based on the prediction result, dynamically adjusting the temperature and oxygen supply in the first fermentation module using the control system: When g(x) = -1, calculate the adjusted parameter increment Δx 温度 and Δx 氧气浓度 ; Adjusting the temperature to x by means of a temperature control device 温度 + Δx 温度 ; Adjusting the oxygen concentration to x by the oxygen supply system 氧气浓度 + Δx 氧气浓度 ; S26, repeating steps S21 to S25 until the support vector machine prediction model outputs g(x) = +1, completing the generation of the first fermentation liquor.

3. The method of deployment of a mobile fruit wine fermentation plant based on modular integration as claimed in claim 1, wherein, The S3 specifically comprises: S31, introducing the first fermentation liquid into the second fermentation module, dynamically modeling the parameters in the fermentation liquid by using a graph convolution network, and constructing a multi-dimensional relationship graph G=(V, E) of the fermentation process, wherein the node set V represents key fermentation parameters, including temperature, sugar content, alcohol concentration, pH value and oxygen concentration, and the edge set E represents the mutual influence relationship between the parameters, each node v i ∈V has an initial feature vector h i The feature matrix of each node is updated through multi-layer propagation of the graph convolution network: where H (l+1) represents the feature matrix of the l+1th layer, H (l) represents the feature matrix of the lth layer, and σ represents an activation function, represents the degree matrix of the graph adjacency matrix with self-loops, W (l) represents the trainable weight matrix of the lth layer, and α k represents a weight coefficient, and R k represents the matrix of the kth relationship type, and K represents the total number of relationship type matrices; S32, using a generative adversarial network to simulate the nonlinear dynamic changes of sugar conversion to alcohol during fermentation, the generative adversarial network comprising a generator G and a discriminator D, wherein the generator G is composed of a multi-layer adaptive normalization and residual network, and is used to generate a fermentation parameter sequence G(z): G(z) = ResNet(AdaIN(z, S style )); where z denotes an input random noise vector, S style denotes the fermentation condition pattern vector extracted by the style encoder, AdaIN denotes adaptive normalization, and ResNet denotes a residual network. S33, the discriminator D adopts a multi-scale convolution network to distinguish the actual fermentation parameter sequence x and the generated fermentation parameter sequence G(z) from different scales, and optimizes the objective function; S34, after the training is completed, the generator G generates an optimal fermentation parameter sequence, and combines the multi-dimensional feature association output by the graph convolution network to obtain an optimal parameter adjustment vector ΔP = [ΔT, ΔS, ΔO] of each time step: ΔP t = G(z) + GCN(H (l) ). where ΔP t denotes the optimal parameter adjustment vector of the t-th time step, ΔT denotes the temperature adjustment value, ΔS denotes the stirring speed adjustment value, ΔO denotes the oxygen concentration adjustment value, and GCN denotes the graph convolution network. S35, according to the optimal parameter adjustment vector, the control system of the second fermentation module is used for real-time adjustment: Adjust temperature to T t + ΔT; Adjust the stirring speed to S t + ΔS; Adjust the oxygen concentration to O t + ΔO; S36, steps S31 to S35 are repeatedly executed to continuously optimize the fermentation process of the second fermentation module until the second fermentation liquor is output.

4. The method of deployment of a mobile fruit wine fermentation plant based on modular integration according to claim 1, characterized in that, The S4 specifically comprises: S41, introducing the second fermentation liquor into a third fermentation module, modeling the space-time dynamic characteristics of the flavor substances in the fermentation liquor by using a space-time graph neural network, and constructing a multi-dimensional space-time feature map G of the fermentation liquor s = (V s , E s ), wherein V s represents a key flavor substance parameter node set at time s, including acidity, ester and phenolic substance concentration, E s represents the space-time correlation between parameters, and the node feature matrix is updated through multi-layer propagation of the space-time graph neural network wherein, represents the feature matrix of the q+1th layer at time s, σ represents an activation function, N(j) represents a set of neighbor nodes of node j, η jk represents the spatio-temporal weight between node j and node k, P (q) represents the trainable weight matrix of the spatial convolution kernel, represents the trainable weight matrix of the temporal convolution kernel, Q r represents the feature transfer matrix, δ r represents the feature transfer matrix Q r the influence weight on the flavor substance, represents the feature matrix of the qth layer at time s, R represents the number of feature transfer matrices; S42, dimensionality reduction and reconstruction of the feature space of flavor substances are performed by using a variational autoencoder, the variational autoencoder comprises an encoder and a decoder, the encoder maps the input high-dimensional feature vector Z s into a latent space U, and the decoder reconstructs the latent variable u into the original feature space, and the latent space distribution of the encoder is: p ψ (u|Z s )=N(u|μ(Z s ),Σ(Z s )); where p ψ (u | Z s ) denotes the probability distribution of the latent variable u given the input Z s , μ(Z s ) denotes the mean of the latent variable u, Σ(Z s ) denotes the covariance matrix of the latent variable u, and N(u | μ(Z s ), Σ(Z s )) denotes a normal distribution. S43, sampling the latent variables by variational inference to obtain the optimal feature expression of flavor substances, combining the spatio-temporal dynamic modeling of parameters by the spatio-temporal graph neural network to obtain the optimal aging conditions, including the adjustment parameters of the micro-oxygen environment and the temperature, and constructing the optimal parameter adjustment vector ΔV s =[ΔO, ΔR], wherein: ΔO represents a micro-oxygen environment adjustment amount, which is used to adjust the oxygen concentration in the fermentation liquor; ΔR represents a temperature adjustment amount, which is used to adjust the temperature in the fermentation liquor; S44, according to the optimal parameter adjustment vector, the control system of the third fermentation module is used for real-time adjustment: Adjusting the micro-oxygen environment to O s + ΔO, wherein O s represents the oxygen concentration at time s; Adjust temperature to R s + ΔR, where R s represents the temperature at time s; S45, steps S41 to S44 are repeatedly executed to continuously optimize the aging process of the third fermentation module until the flavor optimization and clarification treatment are completed, and the finished fruit wine is obtained.

5. The method of deployment of a mobile fruit wine fermentation plant based on modular integration as claimed in claim 1, wherein, The S6 specifically comprises: S61, transmit the real-time data collected by the sensor network in each fermentation module to the central control system to construct a global data set D g ={X1,X2,X3}, wherein X1 represents the real-time monitoring data of the first fermentation module, X2 represents the real-time monitoring data of the second fermentation module, and X3 represents the real-time monitoring data of the third fermentation module; S62, using a multi-layer coordination algorithm on the global data set D g analysis, using a method combining reinforcement learning and Bayesian optimization, defining the state space S t , action space A t and reward function R t ; S63, in the reinforcement learning framework, the parameter adjustment strategy is learned and optimized by using a deep Q network, and the goal is to maximize the cumulative reward function R t : where Q * (S t , A t ) represents the optimal Q-value of performing action A t at state S t , π represents the probability of selecting an action at each state, and γ represents a discount factor; S64, the parameters θ of each fermentation module are further fine-tuned by using Bayesian optimization, a Gaussian process model is used to approximate the objective function, and the next evaluation point is selected by maximizing the expected improvement value: where θ t+1 represents the next parameter configuration, and EI(θ) represents the expected improvement value; S65, the optimal parameter adjustment scheme obtained by reinforcement learning and Bayesian optimization is fed back to each fermentation module, and the fermentation conditions of each module are adjusted in real time; S66, steps S61 to S65 are repeatedly executed to continuously coordinate and optimize the entire fermentation process.

6. The method of deployment of a mobile fruit wine fermentation plant based on modular integration according to claim 5, characterized in that, The S62 specifically comprises: S621, State Space S t It consists of the current parameter state of each fermentation module, denoted as S. t = {s1, s2, s3}, where s i This indicates the parameter status for the i-th fermentation module; S622, Action space A t The parameter adjustment operations executable by the central control system constitute an action space A t = {a1, a2, a3}, where a i denotes the parameter adjustment operation performed on the i-th fermentation module; S623, reward function R t defined as a comprehensive evaluation function of the sugar conversion efficiency, alcohol production rate, and flavor formation degree in the fermentation process: R t = w1E(sugar conversion) + w2E(alcohol generation) + w3E(flavor balance); Wherein, w1, w2 and w3 represent weight coefficients, E(sugar conversion) represents the efficiency of sugar to alcohol conversion, E(alcohol generation) represents the alcohol generation rate, and E(flavor balance) represents the balance degree of flavor substances; wherein represents the instantaneous alcohol production rate, represents the instantaneous by-product production rate, C 糖 (t) represents the sugar concentration at time t, represents the rate of change of acidity, p1and p2represent weight coefficients, τ represents a tuning coefficient, ε represents a smoothing term, t1represents the end time of the integration, t0represents the start time of the integration; where M represents the number of observation samples, θ0represents a constant term of the model, ΔC 酒精 (t-p) represents the change in alcohol concentration at time t-p, θ p represents an autoregressive coefficient of the autoregressive moving average model, φ q represents a moving average coefficient of the autoregressive moving average model, ε t-q represents a random disturbance term at time t-p, P represents the order of the autoregressive model part, Q represents the order of the moving average model part, δ represents a coefficient of the adjustment strength of the function f(T, O2, S) to the alcohol generation rate, T represents the temperature, O2represents the oxygen concentration, and S represents the stirring rate: f(T, 02, S) = a T tanh(T) + a O2 log2(l + 02) + a S sin(S); wherein α T , α O2 , and α S represent influence coefficients; where exp represents an exponential function, H(C 风味 ) represents the complexity of the flavor substance, ξ i represents a weight coefficient, and I(C 风味,i ; C 目标风味,i ) represents the mutual information between the current concentration of the i-th flavor substance and the target flavor concentration. where p(Ci) represents the distribution probability of the concentration of the i-th flavor substance; and 风味,i ) represents the distribution probability of the concentration of the i-th flavor substance; and where p(c i ,c j ) represents the joint probability distribution of the i-th flavor substance and the j-th target flavor, p(c i ) represents the marginal probability distribution of the i-th flavor substance, and p(c j ) represents the marginal probability distribution of the j-th target flavor.

Images